Photovoltaic output prediction method and system combined with ground-based cloud picture

Abstract

The invention discloses a photovoltaic output prediction method and system combined with a foundation cloud atlas, and the method comprises the steps: extracting the global features and local features of a cloud layer based on the foundation cloud atlas, and inputting a cloud atlas prediction model to carry out the prediction of environment parameters; correcting the environmental parameters predicted by the numerical weather forecast and the environmental parameters predicted by the cloud picture prediction model by using the actually measured environmental parameters through a variance weight distribution method to obtain corrected environmental parameters; establishing a slope irradiance mechanism model, and calculating slope irradiance according to the corrected irradiance data; and inputting the slope irradiance and the corrected environmental parameters into a trained photovoltaic output prediction model, and outputting a photovoltaic output prediction result. According to the invention, by fusing the global features and the local features of the ground-based cloud atlas, the limitation of traditional single meteorological data prediction is broken through; through the combination of the mechanism and the data, the overfitting risk caused by independently using a neural network to carry out data driving is avoided, and the accuracy of final photovoltaic output prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic output technology, and in particular to a photovoltaic output prediction method and system combined with a foundation cloud map. Background Art

[0002] In recent years, global fossil fuel shortages and environmental pollution have attracted widespread attention. The global use of renewable energy has continued to advance, and photovoltaic power generation has entered a stage of large-scale development. Photovoltaic power generation prediction technology, as a key support for the efficient operation of photovoltaic power generation systems, has become increasingly important. However, photovoltaic power generation output is affected by many factors, especially meteorological conditions such as solar radiation intensity, cloud cover, and temperature, which make its output highly intermittent, volatile, and uncertain. As the proportion of photovoltaic power generation in the energy structure continues to increase, the integration of large-scale photovoltaic power generation into the grid poses challenges to the safe and stable operation of the power system. Accurately predicting photovoltaic power generation power is an important means to improve photovoltaic utilization and ensure the safe and stable operation of the power system. Therefore, there is an urgent need for a photovoltaic output prediction method that can fully consider the changes in meteorological factors that affect photovoltaic power generation output and accurately predict photovoltaic power generation power.

[0003] A "Photovoltaic Output Forecasting Method" disclosed in Chinese patent literature, with publication number CN113177652A and publication date July 27, 2021, involves: building a cloud-edge collaborative framework; establishing an initial prediction module using a Bayesian recurrent neural network in a cloud computing center; using the prediction module to predict photovoltaic output based on real-time data and uploading it to edge computing devices; and adjusting parameters and updating the prediction module based on prediction deviations in the cloud computing center. This technology incorporates Bayesian principles and uses the Monte Carlo dropout method for approximate inference. Network weights are learned by minimizing the KL divergence between the approximate distribution of network weights and the posterior distribution. Prediction deviations are calculated using the squared Mahalanobis distance or local density ratio to adjust network weights. This technology improves prediction accuracy and robustness, providing effective scheduling reference information for photovoltaic power plant operations and improving economic efficiency. However, the specific technical content primarily relies on extensive historical photovoltaic power generation output data for prediction model training and photovoltaic output forecasting. It does not specifically consider factors influencing photovoltaic output, particularly the impact of cloud cover on solar irradiance fluctuations, leading to significant room for improvement in prediction accuracy. Summary of the Invention

[0004] The present invention aims to overcome the problem in the prior art that the photovoltaic output fluctuations caused by the impact of actual cloud factors on solar irradiation changes are not given much consideration, resulting in insufficient accuracy of the final photovoltaic output prediction results. A photovoltaic output prediction method and system combined with ground-based cloud maps are provided.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A photovoltaic output prediction method combined with a ground-based cloud map includes: The global and local features of the cloud layer are extracted based on the ground-based cloud image and input into the cloud image prediction model to predict the environmental parameters. The environmental parameters predicted by the numerical weather forecast and the environmental parameters predicted by the cloud image prediction model are corrected using the measured environmental parameters through the variance weight distribution method to obtain the corrected environmental parameters. Establish a slope irradiance mechanism model and calculate the slope irradiance with the corrected irradiance data; The slope irradiance and the corrected environmental parameters are input into the trained photovoltaic output prediction model to output the photovoltaic output prediction results.

[0006] The present invention proposes a photovoltaic output prediction method based on mechanism and data fusion to more accurately predict the future power generation capacity of photovoltaic systems. By integrating the global and local characteristics of ground-based cloud maps, the limitations of traditional single meteorological data prediction are broken through, and the prediction accuracy of environmental parameters such as irradiance, temperature and wind speed is significantly improved; the variance weight distribution method is adopted to correct the environmental parameters predicted by numerical weather forecasts and ground-based cloud maps with measured environmental parameters to obtain more accurate environmental parameters; the actual physical laws are constrained by the slope irradiance mechanism model, and the correlation between historical photovoltaic output and multiple environmental parameters is explored in combination with the neural network model, avoiding the risk of overfitting when using neural networks alone for data-driven, thereby improving the accuracy of the final photovoltaic output prediction.

[0007] Preferably, the cloud image prediction model performs mapping of ground-based cloud images to environmental parameters; The global features of the cloud layer extracted based on the ground-based cloud image include: The red-to-blue pixel ratio in the RGB channel of the ground-based cloud image is calculated to generate a cloud thickness matrix that distinguishes cloud types. The cloud thickness matrix is used to calculate the proportion of cloud pixels to obtain the cloud cover. The cloud thickness matrix is converted into a cloud cover matrix, and the cloud thickness coefficient is calculated in combination with the brightness matrix of the ground-based cloud image.

[0008] Preferably, the correction using the measured environmental parameters through the variance weight distribution method includes: The variance of the error between the environmental parameters predicted by numerical weather forecast and the measured environmental parameters is taken as the first variance; The variance of the error between the environmental parameters predicted by the cloud image prediction model and the measured environmental parameters is used as the second variance; According to the principle of minimum overall variance, weights are assigned to the environmental parameters predicted by the two methods respectively, and the weighted sum is taken to obtain the corrected environmental parameters.

[0009] Preferably, the calculation of the cloud coverage ratio includes: the cloud thickness matrix represents the cloud category corresponding to each pixel point in the ground-based cloud map; The cloud cover rate is obtained by taking the ratio of the number of cloud pixels, including thin clouds and thick clouds, to the total number of pixels in the cloud thickness matrix.

[0010] Preferably, the calculation of the cloud thickness coefficient includes: Assign corresponding weight coefficients to different cloud categories in the cloud thickness matrix to obtain the cloud cover matrix; Use the HSV color model to process the ground-based cloud image and extract the value representing the image brightness to construct the brightness matrix; The cloud thickness coefficient is obtained by taking the ratio of the dot product of the cloud cover matrix and the brightness matrix to the sum of all elements in the brightness matrix.

[0011] Preferably, the variance calculation of the numerical weather forecast or cloud image prediction model includes: each environmental parameter includes a number of data samples for variance calculation; Calculate the absolute percentage error between the predicted environmental parameters and the measured environmental parameters for each sample, as well as the average absolute percentage error of all samples; The variance is obtained by summing the squares of the differences between each absolute percentage error and the mean absolute percentage error and dividing by the number of samples.

[0012] Preferably, the weighting of the environmental parameters predicted by the two methods includes: The sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the first variance is used as the numerator to obtain the weight of the numerical weather forecast prediction; the sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the second variance is used as the numerator to obtain the weight of the cloud map prediction model prediction.

[0013] Preferably, the slope irradiance mechanism model is: Slope irradiance includes direct irradiance, scattered irradiance and reflected irradiance received by the slope; The scattered irradiance received by the inclined surface is the product of the scattered irradiance of the horizontal surface and the scattering coefficient. The scattering coefficient is the sum of the cosine of the photovoltaic panel inclination angle and one divided by two. The reflected irradiance received by the inclined surface is the product of the total irradiance of the horizontal surface and the ground reflectivity and reflection coefficient. The reflection coefficient is the difference between one and the cosine of the inclination angle of the photovoltaic panel divided by two.

[0014] Preferably, the generation of the cloud thickness matrix comprises: According to the solar zenith angle of the ground-based cloud image, the corresponding latest cloud image is selected from the clear sky library, and the atmospheric correction factor is calculated using the latest cloud image. The difference between the red-to-blue pixel ratio of the ground-based cloud image and the red-to-blue pixel ratio of the latest cloud image is calculated, and the cloud category of each pixel is judged by combining the atmospheric correction factor and the preset cloud category threshold. The cloud category of each pixel is used to form a cloud thickness matrix.

[0015] A photovoltaic output prediction system combined with a ground-based cloud map, comprising: Numerical weather forecast module, which performs numerical weather forecast of environmental parameters and saves corresponding historical data; The cloud map prediction model module extracts the global and local features of the ground-based cloud map, performs cloud map prediction model predictions on environmental parameters, and saves the corresponding historical data; Environmental parameter correction module, which corrects the predicted results of environmental parameters according to the measured environmental parameters; The photovoltaic output prediction module stores the slope irradiance mechanism model and the photovoltaic output prediction model to perform photovoltaic output prediction.

[0016] The present invention has the following beneficial effects: by integrating the global and local features of the ground-based cloud map, it breaks through the limitations of traditional single meteorological data prediction and significantly improves the prediction accuracy of environmental parameters such as irradiance, temperature and wind speed; adopts the variance weight distribution method to correct the environmental parameters of numerical weather forecasts and ground-based cloud map predictions with measured environmental parameters to obtain more accurate environmental parameters; constrains the actual physical laws through the slope irradiance mechanism model, combines the neural network model to explore the correlation between historical photovoltaic output and multiple environmental parameters, avoids the risk of overfitting when using neural networks alone for data-driven, and improves the accuracy of the final photovoltaic output prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the photovoltaic output prediction method in the present invention.

[0018] Figure 2 This is a flow chart of the present invention for making corrections using the variance weight distribution method using the measured environmental parameters. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] Existing photovoltaic output prediction methods include those based on physical models, statistical analysis, and machine learning. Among them, physical model-based prediction methods include numerical weather forecasts; however, they rely heavily on updated numerical weather forecast data and cannot capture meteorological changes such as sudden changes in cloud cover over a short period of time. Statistical analysis-based prediction methods include time series analysis based on historical output data, capturing the periodicity of output for prediction; however, they cannot effectively handle the nonlinear relationship between irradiance and environmental parameters and photovoltaic output, resulting in large fluctuations in prediction errors. Machine learning-based prediction methods include using historical data to train models and learn the mapping relationship between environmental parameters and photovoltaic output for prediction; however, their selection features are relatively simple, typically using only conventional parameters such as temperature and irradiance, ignoring cloud characteristics, and easily ignoring physical constraints when driven purely by data, leading to overfitting problems.

[0021] To solve the above problems, the present invention provides Figure 1 A photovoltaic output prediction method combined with a ground-based cloud map is shown, including: The global and local features of the cloud layer are extracted based on the ground-based cloud image and input into the cloud image prediction model to predict the environmental parameters. The environmental parameters predicted by the numerical weather forecast and the environmental parameters predicted by the cloud image prediction model are corrected using the measured environmental parameters through the variance weight distribution method to obtain the corrected environmental parameters. Establish a slope irradiance mechanism model and calculate the slope irradiance with the corrected irradiance data; The slope irradiance and the corrected environmental parameters are input into the trained photovoltaic output prediction model to output the photovoltaic output prediction results.

[0022] It should be noted that the present invention proposes a photovoltaic output prediction method based on mechanism and data fusion to more accurately predict the power generation capacity of future photovoltaic systems. By integrating the global and local characteristics of ground-based cloud maps, it breaks through the limitations of traditional single meteorological data prediction and significantly improves the prediction accuracy of environmental parameters such as irradiance, temperature and wind speed. The variance weight distribution method is used to correct the environmental parameters of numerical weather forecasts and ground-based cloud map predictions with measured environmental parameters to obtain more accurate environmental parameters. The actual physical laws are constrained by the slope irradiance mechanism model, and the neural network model is combined to explore the correlation between historical photovoltaic output and multiple environmental parameters, avoiding the risk of overfitting when using neural networks alone for data-driven, thereby improving the accuracy of the final photovoltaic output prediction.

[0023] It is worth noting that the environmental parameters in this invention are a general term for various environmental factors in the space surrounding photovoltaic power generation modules that affect power generation output. These include parameters such as ambient temperature, humidity, wind speed, wind direction, and irradiance, which can fluctuate in the short term, leading to changes in photovoltaic power generation output. Irradiance includes direct irradiance, diffuse irradiance, and total irradiance. The subsequent prediction of environmental parameters involves predicting each of the influencing factors contained therein. Each factor is predicted and corrected to obtain a corrected prediction result, which is then substituted into the slope irradiance mechanism model and photovoltaic output prediction model to accurately predict photovoltaic output.

[0024] Specifically, various data collection and preprocessing are first carried out before photovoltaic output prediction. Various measured data of the target photovoltaic power station are collected, including the environmental parameters, geographical parameters, equipment parameters, historical output parameters, etc. of the photovoltaic power station, including ambient temperature, humidity, wind speed, wind direction, scattered irradiance, direct irradiance, total irradiance, longitude and latitude, altitude, component current, voltage, historical output, foundation cloud map, etc.

[0025] Based on the geographic location of the target PV power station, historical numerical weather forecast data is obtained, including diffuse irradiance, direct irradiance, total irradiance, temperature, humidity, and wind speed. Data is cleaned, including outlier removal, missing value filling, and data formatting. Specified parameter ranges are determined, and data outside these ranges are identified as outliers and removed. Missing values are filled with the average value of the previous and next time points to ensure data accuracy and completeness, simplify subsequent data processing, and improve analytical accuracy.

[0026] After data preprocessing, multiple features of the ground-based cloud image are extracted to train the cloud image prediction model for environmental parameter prediction. Intelligent algorithms such as convolutional neural networks are used to capture more spatial information and key features, such as large-scale cloud structures, to extract local features from the ground-based cloud image. Local features of the ground-based cloud image are those related to cloud edges or cloud structure. This is achieved by training the local feature extraction model on a large number of annotated sample images as a training set. Since this technical aspect is a common technique, it will not be described in detail.

[0027] In addition to local feature extraction, it is also necessary to combine global feature parameters such as cloud coverage and cloud thickness coefficient extracted by knowledge-based cloud detection technology as input to the cloud image prediction model; train the model to establish an accurate mapping relationship between ground-based cloud images and key environmental parameters such as irradiance (including diffuse irradiance, direct irradiance and total irradiance) / ambient temperature / wind speed, so as to achieve accurate prediction and estimation from ground-based cloud images to environmental parameters.

[0028] Predicting environmental parameters based on cloud-atlas prediction models is a method for obtaining environmental parameters, and is suitable for situations where cloud-related factors have a strong influence. Furthermore, numerical weather forecasts are required to predict the same environmental parameters, which is suitable for accurately predicting PV output in clear, light-cloudy weather. Using a weighted allocation method, the environmental parameter prediction results of the two prediction methods are assigned corresponding weights using measured environmental parameters. The weighted summation of the two prediction results yields a corrected, more accurate environmental parameter prediction.

[0029] Corrected environmental parameters include corrected irradiance, temperature, wind speed, and humidity. Based on the principles of photovoltaic energy conversion and combined with predicted data for corrected irradiance, ambient temperature, wind speed, and other environmental parameters, a knowledge-data-fused photovoltaic output prediction model was developed to accurately predict ultra-short-term photovoltaic output. The collected PV power station location information, performance parameters, and environmental parameters such as corrected irradiance were incorporated into the slope irradiance mechanism model to calculate the slope irradiance. Using an LSTM neural network, the model was trained using slope irradiance, corrected ambient temperature, wind speed, ambient humidity, and photovoltaic module performance data to develop a photovoltaic output prediction model that accurately predicts photovoltaic power generation.

[0030] In addition to providing a photovoltaic output prediction method combined with a foundation cloud map, the present invention also provides a photovoltaic output prediction system combined with a foundation cloud map, including: Numerical weather forecast module, which performs numerical weather forecast of environmental parameters and saves corresponding historical data; The cloud map prediction model module extracts the global and local features of the ground-based cloud map, performs cloud map prediction model predictions on environmental parameters, and saves the corresponding historical data; Environmental parameter correction module, which corrects the predicted results of environmental parameters according to the measured environmental parameters; The photovoltaic output prediction module stores the slope irradiance mechanism model and the photovoltaic output prediction model to perform photovoltaic output prediction.

[0031] It should be noted that, in addition to the aforementioned modules, the present invention also includes a data acquisition and processing module capable of collecting geographic parameters, environmental parameters, equipment parameters, historical output parameters, historical numerical weather forecast data, and historically measured environmental parameters of the photovoltaic power station. The system's modular design supports rapid adaptation to photovoltaic power stations of varying sizes and locations. For example, when connecting a new power station, photovoltaic processing forecasts can be performed by simply updating the geographic parameters and component performance data in each module.

[0032] It is worth noting that the system of the present invention introduces multi-feature extraction from ground-based cloud maps, acquiring cloud dynamic information such as cloud coverage and cloud thickness coefficient in real time based on the cloud maps. This can improve prediction accuracy in scenarios with rapidly changing cloud layers and enhance adaptability to local meteorological changes. For multi-feature extraction of ground-based cloud maps, both global and local features are input into the cloud map prediction model for prediction, enabling simultaneous utilization of the macroscopic coverage and microstructure of the cloud layer to improve the spatiotemporal resolution of irradiance prediction. Furthermore, the prediction results of environmental parameters are corrected based on the measured environmental parameters, adaptively correcting the weights corresponding to different prediction methods to avoid prediction errors caused by the failure of a single model.

[0033] As a specific example, the cloud image prediction model maps ground-based cloud images to environmental parameters. After inputting the global and local features of the ground-based cloud image, it outputs corresponding environmental parameter predictions, including irradiance, temperature, humidity, and wind speed. Based on the local features of the ground-based cloud image, a trained neural network model can be used to directly extract features from the ground-based cloud image. The specific training methods are conventional techniques and are not described in detail.

[0034] The global features of clouds are extracted from ground-based cloud images, including: calculating the red-to-blue pixel ratio in the RGB channel of the ground-based cloud image to generate a cloud thickness matrix CT that distinguishes cloud categories; calculating the proportion of cloud pixels using the cloud thickness matrix CT to obtain the cloud coverage; converting the cloud thickness matrix CT into a cloud coverage matrix A, and calculating the cloud thickness coefficient in combination with the brightness matrix L of the ground-based cloud image.

[0035] It should be noted that the size of the cloud thickness matrix CT is identical to that of the ground-based cloud image. Each element in the cloud thickness matrix CT corresponds to the cloud category of the pixel at the same location in the ground-based cloud image. It primarily distinguishes thin cloud, thick cloud, and clear-sky pixels. The specific cloud thickness matrix CT can be obtained by comparing the difference between the red-to-blue pixel ratio of the ground-based cloud image and the red-to-blue pixel ratio of the clear-sky database with different classification thresholds. Alternatively, the difference between the two can be optimized using an atmospheric correction factor and then compared with different classification thresholds. Once the cloud thickness matrix CT is obtained, the cloud cover fraction and cloud thickness coefficient can be further calculated to obtain the global characteristics of the ground-based cloud image.

[0036] Specifically, the calculation of cloud cover includes: The cloud cover rate is obtained by taking the ratio of the number of cloud pixels (including thin clouds and thick clouds) in the cloud thickness matrix CT to the total number of pixels in the cloud thickness matrix CT.

[0037] Based on the number of thin cloud, thick cloud, and clear sky pixels in the cloud thickness matrix CT, the proportion of cloud pixels (including thin and thick clouds) in the entire image is calculated to obtain the cloud cover fraction c. Here, i and j represent the pixel dimensions of the image, I and J represent the image size, and C1 and C2 are assigned values of 0 or 1, respectively, depending on the pixel classification.

[0038] Specifically, the calculation of cloud thickness coefficient includes: Assign corresponding weight coefficients to different cloud types in the cloud thickness matrix CT to obtain the cloud cover matrix A; Use the HSV color model to process the ground-based cloud image and extract the value representing the image brightness to construct the brightness matrix L; The cloud thickness coefficient is obtained by taking the ratio of the dot product of the cloud cover matrix A and the brightness matrix L to the sum of all elements in the brightness matrix L.

[0039] Where Tr[LA T ] represents the sum of the products of each element in the brightness matrix L and the corresponding element in the cloud cover matrix A; L ij Represents the brightness value at coordinate (i, j) in the brightness matrix L.

[0040] The cloud thickness coefficient can be considered the product of brightness under ideal conditions (accounting for aerosol effects) and cloud cover. First, the cloud image data is processed using the HSV color model to extract values representing image brightness. This value serves as a brightness index and forms the brightness matrix L. For a pixel in the ground-based cloud image, after HSV color model processing, its brightness value V is the maximum of the pixel's RGB values. The size of the brightness matrix L is the same as the pixel size of the ground-based cloud image, and each element of the brightness matrix L represents the brightness value of the pixel at the corresponding location.

[0041] For the cloud thickness matrix CT, different weight coefficients are assigned to the corresponding positions of thin cloud, thick cloud, and clear sky pixels. This matrix is constructed with the same size as the ground-based cloud image (size I*J, where I and J are the pixel sizes of the ground-based cloud image), thereby obtaining the cloud cover matrix A. The cloud thickness coefficient is calculated as the ratio of the sum of the products of each element in the brightness matrix L and the element at the corresponding position in the cloud cover matrix A to the total brightness value of the cloud image. This reflects the overall cloud coverage in the ground-based cloud image.

[0042] As an optional embodiment, the generation of the cloud thickness matrix includes: According to the solar zenith angle of the ground-based cloud image, the corresponding latest cloud image is selected from the clear sky library, and the atmospheric correction factor is calculated using the latest cloud image. The difference between the red-to-blue pixel ratio of the ground-based cloud image and the red-to-blue pixel ratio of the latest cloud image is calculated, and the cloud category of each pixel is judged by combining the atmospheric correction factor and the preset cloud category threshold. The cloud category of each pixel is used to form a cloud thickness matrix.

[0043] It should be noted that in the process of generating the cloud thickness matrix using the red - blue pixel ratio of the ground - based cloud image and the clear - sky library, it is first necessary to establish and update the clear - sky library data in advance to obtain a more accurate cloud thickness matrix. The clear - sky library records historical clear - sky cloud images, their corresponding shooting times, and solar zenith angles. In addition, for the convenience of subsequent calculations, the clear - sky library can also record the red - blue pixel ratios of each pixel point in the clear - sky cloud image. The cloud category thresholds include the clear - sky threshold and the thick - cloud threshold.

[0044] The specific content of establishing the clear - sky library includes: for historical clear - sky cloud images, calculate and record their red - blue pixel ratios. At the same time, the relationships between the red - blue pixel ratio of the clear - sky cloud image, the solar zenith angle SZA, the pixel zenith angle PZA, and the angle SPA between the pixel and the sun can be obtained. The area where PZA is greater than 75° corresponds to the near - horizon area, and the red - blue pixel ratio value is relatively large. The area where SPA is less than 35° corresponds to the solar circle area, and the red - blue pixel ratio value is the largest. This is because in the near - horizon area, the aerosol concentration is high and the transmission path in the atmosphere is long. Near the solar circle, due to the forward scattering of sunlight, the near - horizon and solar - circle areas in the cloud image appear whiter and brighter, and the red - blue pixel ratio values are also larger than those in other areas.

[0045] Calculate the red - blue pixel ratio RBR(i, j) of the ground - based cloud image. Using the solar zenith angle SZA corresponding to the ground - based cloud image as an index, search for the corresponding clear - sky cloud image in the clear - sky library, and select the clear - sky cloud image with the latest shooting time, that is, the latest cloud image, to obtain its red - blue pixel ratio CSL(m, n). The position of (i, j) in the ground - based cloud image corresponds to the position of (m, n) in the latest cloud image.

[0046] For the atmospheric correction factor TCF, calculate the difference in the red - blue pixel ratio Diff(i, j)=RBR(i, j)-CSL(m, n) between the ground - based cloud image and the latest cloud image one by one, and screen out the clear - sky pixel points in the ground - based cloud image according to the relationship between Diff(i, j) and the clear - sky threshold; calculate the ratio between the average value of the red - blue pixel ratios of all clear - sky pixel points and the average value of the red - blue pixel ratios of the corresponding pixel points in the latest cloud image, and use this ratio as the initial atmospheric correction factor TCF′; calculate the absolute error MAD between CSL(m, n) and CSL(m, n)*TCF′; when MAD < 0.05, judge whether 0.8 < TCF′ < 1.2 holds. If it holds, then TCF = TCF′, otherwise, TCF = 1; when MAD ≥ 0.05, assign CSL(m, n)*TCF′ to CSL(m, n) and return the latest cloud image to repeat the above steps until MAD < 0.05. MAD < 0.05 means that the difference in the red - blue pixel ratios between the clear - sky pixel points in the ground - based cloud image and the pixel points in the clear - sky library is not significant.

[0047] Traverse the red-to-blue pixel ratio (RBR(i,j)) of the ground-based cloud image and the corresponding red-to-blue pixel ratio (CSL(m,n)) in the latest cloud image, and calculate DiffTCF(i,j) = RBR(i,j) - CSLTCF(m,n). Here, CSLTCF(m,n) = CSL(m,n) * TCF. Based on the relationship between DiffTCF(i,j) and the cloud classification threshold, identify the cloud type corresponding to each pixel (i,j) on the ground-based cloud image. Specifically, when DiffTCF(i,j) ≤ the clear sky threshold, the pixel is considered clear; when DiffTCF(i,j) ≥ the thick cloud threshold, the pixel is considered thick cloud; and when DiffTCF(i,j) is between the clear sky threshold and the thick cloud threshold, the pixel is considered thin cloud.

[0048] As a specific example, Figure 2 The figure shows a flow chart of correcting environmental parameters by using the measured environmental parameters through the variance weight distribution method, including: The variance of the error between the environmental parameters predicted by numerical weather forecast and the measured environmental parameters is taken as the first variance; The variance of the error between the environmental parameters predicted by the cloud image prediction model and the measured environmental parameters is used as the second variance; According to the principle of minimum overall variance, weights are assigned to the environmental parameters predicted by the two methods respectively, and the weighted sum is taken to obtain the corrected environmental parameters.

[0049] It should be noted that the present invention uses ground-based cloud maps to predict key environmental parameters, while also using numerical weather forecasts (NWPs). However, both parameter prediction methods have certain errors. Therefore, the environmental parameters predicted by both methods are corrected based on the key environmental parameters obtained from actual environmental measurements, achieving a more accurate prediction of environmental parameters.

[0050] It is worth noting that the environmental parameters predicted by the numerical weather forecast and the cloud map prediction model are weighted and then combined to obtain a more accurate environmental parameter prediction result. In cloudy weather, the environmental parameters predicted by the numerical weather forecast have large errors and large fluctuations compared to the measured environmental parameters, while the environmental parameters predicted by the cloud map prediction model are closer to the measured environmental parameters. Therefore, dynamic weighting is assigned, with a smaller weight given to the one with larger variance (the numerical weather forecast) and a larger weight given to the one with smaller variance (the cloud map prediction model). Similarly, in clear skies, the numerical weather forecast is given a larger weight, while the cloud map prediction model is given a smaller weight. The influence of the two prediction methods on the final results can be dynamically considered based on the actual changes in the cloud layer to obtain corrected environmental parameters.

[0051] Specifically, based on the measured parameter values of environmental parameters such as irradiance / ambient temperature / wind speed, the variance weight distribution method is used to correct the environmental parameters such as irradiance / ambient temperature / wind speed predicted by numerical weather forecasts and multi-feature extraction of ground-based cloud images, and an environmental parameter prediction model is constructed to generate prediction data of environmental parameters such as irradiance / ambient temperature / wind speed.

[0052] The estimation data is corrected by using a variance weight dynamic allocation method, including: each environmental parameter includes a number of data samples for variance calculation; Calculate the absolute percentage error between the predicted environmental parameters and the measured environmental parameters for each sample, as well as the average absolute percentage error of all samples; sum the squares of the differences between each absolute percentage error and the average absolute percentage error, and divide the sum by the number of samples to obtain the variance result.

[0053] Among them, J is the number of test samples, e1, e2…e J is the absolute percentage error for each test sample, is the mean absolute percentage error of J test samples. m is the variance of model m, ω m is the weight of model m, and m is 1 and 2, corresponding to the two prediction methods of numerical weather forecast and cloud image prediction model respectively.

[0054] Furthermore, weight assignment of environmental parameters predicted by the two methods includes: The sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the first variance is used as the numerator to obtain the weight of the numerical weather forecast prediction; the sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the second variance is used as the numerator to obtain the weight of the cloud map prediction model prediction.

[0055] f=ω1f1+ω2f2 Among them, f1 and f2 correspond to the environmental parameters predicted by the numerical weather forecast and the environmental parameters predicted by the cloud map prediction model, respectively, and f is the final corrected environmental parameter.

[0056] The two prediction methods are weighted based on the principle of minimizing overall variance. First, the inverse of each variance is taken, the sum of all inverses is calculated, and finally, each inverse is divided by the sum of their inverses to obtain the weight of each prediction method. After the weights are calculated, they are multiplied by the corresponding prediction results to obtain the corrected prediction data for environmental parameters such as irradiance, ambient temperature, and wind speed. By correcting the irradiance, ambient temperature, and wind speed estimated by numerical weather forecasts and ground-based cloud image feature extraction based on actual environmental measurements, accurate predictions of these parameters are achieved, thereby improving the accuracy of photovoltaic power forecasts.

[0057] As a specific embodiment, based on the principle of photovoltaic energy conversion and combined with prediction data of environmental parameters such as irradiance / ambient temperature / wind speed, a photovoltaic output prediction model that integrates knowledge and data is developed to achieve accurate prediction of ultra-short-term photovoltaic output.

[0058] The slope irradiance mechanism model is as follows: the slope irradiance includes the direct irradiance, scattered irradiance and reflected irradiance received by the slope; according to the mechanism of irradiance absorption by photovoltaic modules, a correction model is established for the irradiance absorbed by photovoltaic modules to convert solar radiation into the slope irradiance actually received by photovoltaic modules. The slope irradiance is calculated by adding the direct irradiance, scattered irradiance and ground reflected irradiance.

[0059] Among them PV T is the slope irradiance, PV b is the direct irradiance on the horizontal plane, PV b ×R b Indicates the direct irradiance on the slope, R b is the ratio of the radiation flux on the inclined plane to the horizontal plane, PV d represents the scattered irradiance on the horizontal plane, β is the tilt angle of the photovoltaic panel, is the diffuse irradiance on the slope, ρ is the ground reflectivity, and PV is the total solar radiation on the horizontal plane.

[0060] The scattered irradiance received by the inclined surface is the product of the scattered irradiance of the horizontal surface and the scattering coefficient. The scattering coefficient is the sum of the cosine value of the photovoltaic panel inclination angle and one divided by two; the reflected irradiance received by the inclined surface is the product of the total irradiance of the horizontal surface, the ground reflectivity and the reflection coefficient. The reflection coefficient is the difference between one and the cosine value of the photovoltaic panel inclination angle divided by two.

[0061] The PV power station's geographic location information and performance parameters, obtained through data collection and preprocessing, along with environmental parameters such as the predicted and corrected irradiance, are incorporated into the established slope irradiance mechanism model to calculate the slope irradiance. Using an LSTM neural network, the model is trained using slope irradiance, corrected ambient temperature, wind speed, humidity, and PV module performance data to develop a PV power prediction model, enabling accurate prediction of PV power generation.

[0062] The above embodiments are further elaborations and illustrations of the present invention for ease of understanding, and are not intended to limit the present invention in any way. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A photovoltaic output prediction method combined with foundation cloud map, characterized in that: include: Extract global and local features of clouds based on ground-based cloud images, and input them into the cloud image prediction model to predict environmental parameters; For the environmental parameters predicted by numerical weather forecast and cloud image prediction model, the measured environmental parameters are corrected by variance weight allocation method to obtain the corrected environmental parameters; Establish a slope irradiance mechanism model and calculate the slope irradiance with the corrected irradiance data; The slope irradiance and the corrected environmental parameters are input into the trained photovoltaic output prediction model to output the photovoltaic output prediction results.

2. The photovoltaic output prediction method combined with foundation cloud map according to claim 1 is characterized in that: The cloud image prediction model maps the ground-based cloud image to environmental parameters; The global features of the cloud layer extracted based on the ground-based cloud image include: The red-to-blue pixel ratio in the RGB channel of the ground-based cloud image is calculated to generate a cloud thickness matrix that distinguishes cloud types. The cloud thickness matrix is used to calculate the proportion of cloud pixels to obtain the cloud cover. The cloud thickness matrix is converted into a cloud cover matrix, and the cloud thickness coefficient is calculated in combination with the brightness matrix of the ground-based cloud image.

3. A photovoltaic output prediction method combined with foundation cloud map according to claim 1 or 2, characterized in that: The correction using the measured environmental parameters through the variance weight distribution method includes: The variance of the error between the environmental parameters predicted by numerical weather forecast and the measured environmental parameters is taken as the first variance; The variance of the error between the environmental parameters predicted by the cloud image prediction model and the measured environmental parameters is used as the second variance; According to the principle of minimum overall variance, weights are assigned to the environmental parameters predicted by the two methods respectively, and the weighted sum is taken to obtain the corrected environmental parameters.

4. The photovoltaic output prediction method combined with foundation cloud map according to claim 2 is characterized in that: The calculation of the cloud coverage rate includes: the cloud thickness matrix represents the cloud category corresponding to each pixel point in the ground-based cloud map; The cloud cover rate is obtained by taking the ratio of the number of cloud pixels, including thin clouds and thick clouds, to the total number of pixels in the cloud thickness matrix.

5. The photovoltaic output prediction method combined with foundation cloud map according to claim 2 or 4, characterized in that: The calculation of the cloud thickness coefficient includes: Assign corresponding weight coefficients to different cloud categories in the cloud thickness matrix to obtain the cloud cover matrix; Use the HSV color model to process the ground-based cloud image and extract the value representing the image brightness to construct the brightness matrix; The cloud thickness coefficient is obtained by taking the ratio of the dot product of the cloud cover matrix and the brightness matrix to the sum of all elements in the brightness matrix.

6. The photovoltaic output prediction method combined with foundation cloud map according to claim 3 is characterized in that: The variance calculation of the numerical weather forecast or cloud image prediction model includes: each environmental parameter includes a number of data samples for variance calculation; the absolute percentage error between the predicted environmental parameter and the measured environmental parameter of each sample is calculated, and the average absolute percentage error of all samples is calculated; The variance is obtained by summing the squares of the differences between each absolute percentage error and the mean absolute percentage error and dividing by the number of samples.

7. The photovoltaic output prediction method combined with foundation cloud map according to claim 3 is characterized in that: The weighting of the environmental parameters predicted by the two methods according to the principle of minimum overall variance includes: The sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the first variance is used as the numerator to obtain the weight of the numerical weather forecast prediction; the sum of the reciprocals of the first and second variances is used as the denominator, and the reciprocal of the second variance is used as the numerator to obtain the weight of the cloud map prediction model prediction.

8. The photovoltaic output prediction method combined with foundation cloud map according to claim 1, 2, 4, 6 or 7, characterized in that: The slope irradiance mechanism model is: Slope irradiance includes direct irradiance, scattered irradiance and reflected irradiance received by the slope; The scattered irradiance received by the inclined surface is the product of the scattered irradiance of the horizontal surface and the scattering coefficient. The scattering coefficient is the sum of the cosine of the photovoltaic panel inclination angle and one divided by two. The reflected irradiance received by the inclined surface is the product of the total irradiance of the horizontal surface and the ground reflectivity and reflection coefficient. The reflection coefficient is the difference between one and the cosine of the inclination angle of the photovoltaic panel divided by two.

9. The photovoltaic output prediction method combined with foundation cloud map according to claim 2, characterized in that: The generation of the cloud thickness matrix includes: According to the solar zenith angle of the ground-based cloud image, the corresponding latest cloud image is selected from the clear sky library, and the atmospheric correction factor is calculated using the latest cloud image. The difference between the red-to-blue pixel ratio of the ground-based cloud image and the red-to-blue pixel ratio of the latest cloud image is calculated, and the cloud category of each pixel is judged by combining the atmospheric correction factor and the preset cloud category threshold. The cloud category of each pixel is used to form a cloud thickness matrix.

10. A photovoltaic output prediction system combined with a foundation cloud map, applicable to the photovoltaic output prediction method according to any one of claims 1 to 9, characterized in that: include: Numerical weather forecast module, which performs numerical weather forecast of environmental parameters and saves corresponding historical data; The cloud map prediction model module extracts the global and local features of the ground-based cloud map, performs cloud map prediction model predictions on environmental parameters, and saves the corresponding historical data; Environmental parameter correction module, which corrects the predicted results of environmental parameters according to the measured environmental parameters; The photovoltaic output prediction module stores the slope irradiance mechanism model and the photovoltaic output prediction model to perform photovoltaic output prediction.

Images